Spotlighting on Objects: Prior Knowledge-Driven Maritime Image Dehazing and Object Detection Framework

Abstract

Maritime environments often face visibility challenges due to haze which significantly impacts detection models. However, existing maritime object detection algorithms often neglect haze conditions or the unique characteristics of the maritime environment, resulting in decreased effectiveness in hazy weather. In this article, we propose a prior knowledge-driven maritime image dehazing and object detection framework (MDD), which consists of a detection network and a restoration network. Leveraging the characteristics of the highlighted ships in the inverted dark channel prior (IDCP), the detection network incorporates a prior subnetwork to learn ship-related features, which are subsequently merged into the backbone network through an IDCP cross-attention module. During training, the restoration network is integrated to improve the clarity of the features learned by the detection network. In addition, a ship-haze enrichment strategy is implemented to emphasize ship regions in the training samples, along with a ship-aware reconstruction loss to enhance the network's ability to learn dehazed features. Moreover, we establish a maritime object recognition with haze levels (MORHL) data set to evaluate object detector performance in maritime hazy conditions. It includes 13 280 annotated images across six categories: cargo ship, container ship, fishing boat, passenger ship, island, and buoy, with haze levels categorized as light, medium, and heavy. Comprehensive experiments on the MORHL and SMD data sets demonstrate that the proposed MDD framework outperforms the state-of-the-art detectors and various combinations of dehazing and detection methods.